MESA: Improving MoE Safety Alignment via Decentralized Expertise
On the map: LLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks
also in 2 more
Abstract (arXiv)
Mixture-of-Experts (MoE) architectures scale Large Language Models (LLMs) efficiently, enabling greater capacity with reduced computational cost by dynamically routing inputs to relevant experts, yet introduce a critical vulnerability: Safety Sparsity, where safety capabilities concentrate in few experts, making them susceptible to adversarial bypassing. Meanwhile, conventional alignment methods uniformly adapt all parameters, ignoring their functional differences and inadvertently degrading performances. To address these challenges, we propose MESA (MoE Safety Alignment), a targeted alignment framework for MoE-based LLMs that strategically decentralizes safety responsibility to maximize coverage while minimizing interference with utility. Based on Optimal Transport (OT) theory, MESA operates through two mechanisms: (1) Expert Capacity Reallocation uses a transport cost matrix to distribute safety duties to the most cost-effective experts, and (2) Dynamic Routing Refinement constrains the router to precisely activate these decentralized modules. Experiments show that MESA achieves robust defensive performance against varied harmful benchmarks while preserving helpfulness. Code is available at https://github.com/lorraine021/MESA.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 39 of its 42 references that are in the map. First posted on arXiv 30 May 2026.
Builds on
39 of 42 references.
Closest ancestors · 6
Same direction, newest first
- 2025Oyster-I: Beyond Refusal -- Constructive Safety Alignment for Responsible Language Models2 Sep
- 2025SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification20 Jun
- 2025Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable1 Mar
- 2025STAIR: Improving Safety Alignment with Introspective Reasoning4 Feb
- 2024Safety Alignment Should Be Made More Than Just a Few Tokens Deep10 Jun
- 2023Safe RLHF: Safe Reinforcement Learning from Human Feedback19 Oct
Other references · 33
- 2023GPT-4 Technical Reportcited by 18,396
- 2022Training language models to follow instructions with human feedbackcited by 15,971
- 2017Proximal Policy Optimization Algorithmscited by 15,215
- 2021Training Verifiers to Solve Math Word Problemscited by 10,109
- 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
- 2025Qwen3 Technical Reportcited by 8,820
- 2021Evaluating Large Language Models Trained on Codecited by 8,320
- 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
- 2020Measuring Massive Multitask Language Understandingcited by 7,963
- 2025Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilitiescited by 3,796
- 2023Let's Verify Step by Stepcited by 3,789
- 2021Program Synthesis with Large Language Modelscited by 3,637
21 more
- 2021Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsitycited by 3,073
- 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
- 2013Sinkhorn Distances: Lightspeed Computation of Optimal Transportation Distancescited by 2,316
- 2023Jailbreaking Black Box Large Language Models in Twenty Queriescited by 1,420
- 2024DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Modelcited by 1,135
- 2024How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMscited by 573
- 2024JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Modelscited by 536
- 2023DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Modelscited by 518
- 2023UltraFeedback: Boosting Language Models with Scaled AI Feedbackcited by 406
- 2024A StrongREJECT for Empty Jailbreakscited by 385
- 2024TrustLLM: Trustworthiness in Large Language Modelscited by 285
- 2024WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Modelscited by 223
- 2024Thinking LLMs: General Instruction Following with Thought Generationcited by 45
- 2024S-Eval: Towards Automated and Comprehensive Safety Evaluation for Large Language Modelscited by 33
- 2025Steering MoE LLMs via Expert (De)Activationcited by 26
- 2025Inverse Reinforcement Learning with Dynamic Reward Scaling for LLM Alignmentcited by 20
- 2025DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarioscited by 19
- 2025Breaking the Ceiling: Exploring the Potential of Jailbreak Attacks through Expanding Strategy Spacecited by 17
- 2025BadMoE: Backdooring Mixture-of-Experts LLMs via Optimizing Routing Triggers and Infecting Dormant Expertscited by 14
- 2025Strata-Sword: A Hierarchical Safety Evaluation towards LLMs based on Reasoning Complexity of Jailbreak Instructionscited by 11
- 2026Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMscited by 9
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026RASA: Routing-Aware Safety Alignment for Mixture-of-Experts Models
- 2026SEAL: Reinforcing Global Safety in Mixture-of-Experts through Shared Expert ALignment
- 2026RouteHijack: Routing-Aware Attack on Mixture-of-Experts LLMs
- 2025Defending MoE LLMs against Harmful Fine-Tuning via Safety Routing Alignment
- 2025GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMs
- 2026Large Language Lobotomy: Jailbreaking Mixture-of-Experts via Expert Silencing
- 2025Exploiting the Experts: Unauthorized Compression in MoE-LLMs
- 2026RASET: Router-Agnostic Safety-Critical Expert Tuning Exposes Localized Safety Enforcement Failures in Mixture-of-Experts LLMs